Sam Bajracharya

Credit Risk Under Regime Change

Python | Statistical Modelling | Time Series Evaluation

This project investigates how credit risk models behave when the underlying environment changes, focusing on the period surrounding the 2008 financial crisis. Using loan-level data from LendingClub (2007–2015 loan dataset), the objective is not only to predict default, but to understand how model outputs degrade when relationships between borrower characteristics and outcomes shift over time.

The dataset includes borrower and loan attributes such as credit grade, interest rate, debt-to-income ratio, income, loan amount, and term. These variables reflect how risk is assessed and priced within the system, linking model predictions to decisions around which loans appear attractive and which represent elevated exposure.

Two model classes were implemented and compared. Logistic regression was used in both regularised and unregularised forms to capture stable, aggregate relationships between features and default risk. Random forest models were used to capture non-linear interactions and local structure in the data, with hyperparameters tuned using Bayesian optimisation and validated through learning curves.

Evaluation was conducted using AUC-ROC, accuracy, and calibration analysis. Rather than relying on a static train-test split, a rolling window framework was used, where models are trained on recent historical data (approximately 30 days) and evaluated on forward horizons of one to five years. This allows performance to be tracked as conditions evolve.

Key components include:

• Designing a time-aware evaluation framework to assess model reliability under distribution shift
• Comparing linear and tree-based models in terms of predictive performance, stability, and calibration
• Evaluating how predicted default probabilities align with realised outcomes over time
• Tracking shifts in feature importance to identify changing drivers of risk
• Assessing the tradeoff between model flexibility and robustness in non-stationary environments

Results show a clear divergence in model behaviour. Random forests achieve strong performance in stable, pre-crisis conditions, capturing complex interactions within the data. However, performance deteriorates significantly, with reduced AUC and instability in predicted probabilities following the 2008 shift, with reduced generalisation and instability in predicted probabilities. Logistic regression exhibits lower peak performance but maintains more consistent behaviour, with smaller degradation and more reliable calibration.

This distinction is important because model outputs are often used as signals to guide decisions. A model that performs well historically but fails under changing conditions can produce misleading signals, leading to overconfidence in patterns that no longer hold. In contrast, a more stable model may provide weaker signals in normal conditions, but remain usable when the environment shifts.

The difference arises from how each model represents structure. Tree-based models learn fine-grained partitions tied to historical patterns, which become misaligned when the regime changes. Logistic regression imposes a global structure, capturing broader relationships such as the link between leverage, pricing, and default risk, making it more robust to shifts in underlying conditions.

Feature importance analysis further shows that key variables such as loan grade and interest rate change in relevance over time. These features are themselves influenced by the environment, meaning that the mapping from inputs to outcomes evolves, limiting the effectiveness of models trained on past data.

The project highlights a central challenge in modelling real-world systems: predictive performance depends not only on model choice, but on the stability of the environment. In settings where relationships shift, robustness and consistency can matter more than maximising accuracy under historical conditions.

Fraud Detection and Network Forensics (Ongoing)

Python | Network Analysis | Behavioural Analytics

Building a fraud detection and transaction network forensics system using the Elliptic Bitcoin dataset, which contains labelled transaction data (licit vs illicit) together with engineered temporal and structural features. The dataset is modelled as a directed transaction graph, where nodes represent entities and edges represent transfers, allowing analysis of how funds propagate through the network over time.

The project combines statistical modelling and network analysis to identify suspicious behaviour at both the node and subgraph level. Using Python with libraries such as pandas and NumPy for data handling, scikit-learn for classification models, and NetworkX for graph construction and analysis, the work focuses on detecting anomalous transaction patterns, clustering behaviour, and structural signatures associated with illicit activity.

Key objectives include:

• Identifying high-risk nodes using supervised models trained on known illicit transactions
• Detecting anomalous patterns through feature distributions, temporal shifts, and network centrality measures
• Analysing transaction flows to uncover coordinated behaviour, layering patterns, and potential laundering structures
• Exploring how fraud manifests within local neighbourhoods of the graph rather than isolated transactions

The expectation is that fraudulent activity will not appear as random noise, but as structured behaviour within the network, such as tightly connected clusters, unusual flow patterns, or nodes with disproportionate influence or connectivity. The project aims to bridge statistical inference with graph-based reasoning to better understand how illicit financial behaviour emerges and propagates.

Preliminary analysis suggests that illicit activity is not uniformly distributed across the network, but instead appears in locally dense regions characterised by repeated interactions and elevated connectivity. Certain nodes exhibit disproportionately high centrality relative to their transaction volume, indicating potential intermediary roles in fund movement. Temporal patterns also show bursts of activity within short intervals, consistent with layering or rapid redistribution behaviour rather than organic transaction flow.

Quantum Computing in Bioinformatics (Internship)

Qiskit | Optimisation | Scientific Computing | Research

Completed a research internship at the Walter and Eliza Hall Institute of Medical Research, focusing on improving the usability and correctness of a quantum protein-folding model implemented in Qiskit. The work addressed technical and structural issues in a codebase handed down across multiple interns, where poor documentation and incorrect parameterisation limited practical use of the model.

The onboarding process was redesigned by replacing an extensive set of academic readings with a structured glossary and streamlined documentation, allowing new contributors to understand key concepts and begin working with the model significantly faster. In parallel, the implementation was refactored through detailed inline commentary, clarifying theoretical background, modelling assumptions, and execution flow directly within the code.

Key components included:

• Designing a simplified onboarding pathway by consolidating essential concepts into a concise reference document, reducing reliance on external academic sources
• Refactoring a Jupyter notebook implementation with structured inline documentation to improve readability and accessibility for users without prior quantum computing experience
• Diagnosing failures in the Hamiltonian formulation, where constraint penalties (e.g. backtracking and long-range folding) were incorrectly scaled, preventing meaningful interaction between amino acids
• Rescaling constraint terms to restore valid interaction dynamics and enable the optimiser to explore physically meaningful folding configurations
• Replacing gradient-based optimisation with a non-derivative method better suited to the discrete and non-linear structure of the problem space
• Testing the corrected model on simulated configurations to confirm functional behaviour while identifying remaining computational inefficiencies
• Migrating the project from a Google Drive notebook to a structured GitHub repository, introducing version control and improving reproducibility
• Producing a comprehensive technical report and handover documentation to enable future contributors to quickly understand and extend the model

The project emphasised practical problem-solving in complex optimisation systems, combining debugging, model correction, and documentation improvements to transform an unusable research prototype into a functional and maintainable framework for future work.

Where2 (Co-Founder & Developer)

React Native | Distributed Systems | Real-Time State Management

Co-founded and co-developing Where2, a mobile application designed to support group coordination through shared itineraries, messaging, and collaborative decision-making. The project originated within the Melbourne Entrepreneurial Centre startup program, where I joined as the primary technical contributor, taking ownership of system design and implementation as the concept progressed from early-stage idea to a working product.

Designed and implemented a multi-user system where plans, locations, and user interactions are synchronised in real time across devices. The application is built in React Native, with a backend architecture using Supabase to manage persistent state, authentication, and real-time updates across distributed users.

The system has evolved into a multi-component application supporting user accounts, following relationships, location-based planning, and real-time updates across shared itineraries. Core functionality includes dynamic itinerary generation, map-based visualisation of locations and posts, and automated handling of metadata such as update timestamps and time zones. These components are integrated through a structured backend architecture, with persistent storage and access control managed through relational database design.

The central technical challenge in the system is maintaining consistency of shared state across multiple users interacting concurrently. Actions such as editing plans, adding locations, or sending messages must propagate in real time while preserving a coherent view of the itinerary for all participants. This requires careful handling of synchronisation, conflict resolution, and access control to ensure that updates remain consistent under concurrent modification.

Key components included:

• Designing a relational database schema to model users, plans, locations, and social relationships, including following systems and access control for shared plans
• Implementing backend services for user identity, authentication, and state management across concurrent users interacting within the same plan
• Handling real-time synchronisation of plan updates, messaging, and location data across multiple clients, ensuring consistency of shared state
• Structuring application logic to manage user interactions such as invitations, following relationships, and collaborative editing of itineraries
• Developing front-end interfaces in React Native based on iterative Figma prototypes, with emphasis on clarity of interaction and usability in multi-user workflows

Development has moved from prototyping toward release, requiring a shift from feature implementation to system reliability. This includes refining service architecture, ensuring consistency of shared state across users, and validating that interactions behave predictably under real usage. The focus is not only on building functionality, but on delivering a system that remains stable as complexity and user interaction scale.

CSIRO Image2Biomass Datathon

Python | Deep Learning | Prediction Under Limited Data

Developed an end-to-end prediction pipeline to estimate pasture biomass from high-resolution aerial imagery as part of the CSIRO Image2Biomass datathon. The task involved translating large, unstructured visual data into quantitative estimates across multiple targets, including green, dry, and dead plant matter.

The primary challenge was working under limited labelled data and high variability in visual input. Images differed in scale, lighting, and spatial composition, requiring careful preprocessing to preserve meaningful structure while ensuring compatibility with model inputs.

A preprocessing pipeline was implemented to convert wide-format aerial images into consistent square inputs, applying normalisation and standardisation to stabilise feature extraction. A self-supervised Vision Transformer (DINO) was used as a feature extractor, with representations adapted for multi-output regression using PyTorch and optimised via mean squared error loss. This approach leveraged pre-trained representations to improve performance under constrained data conditions.

Key components included:

• Designing a preprocessing pipeline to preserve spatial and textural information relevant to biomass estimation
• Training and evaluating regression models across multiple data splits to assess generalisation
• Combining model outputs using weighted averaging across data splits to improve robustness and reduce overfitting
• Applying post-processing adjustments to correct systematic prediction bias

The project emphasised building a reliable prediction system under real-world constraints. Model performance depended not only on architecture, but on handling noisy inputs, limited labels, and variability in the data distribution. The final pipeline prioritised consistency and robustness over maximising performance on individual splits.

Flood Simulation

QGIS | Spatial Modelling | Risk Analysis

Developed a geospatial flood modelling framework to analyse inundation patterns and evaluate mitigation strategies across flood-prone regions, including areas in Queensland and St Kilda. The project treats flooding as a dynamic system influenced by terrain, water flow, and environmental conditions, rather than a static event.

Raster-based models were used to simulate flood propagation under varying severity scenarios, with interventions such as levees, temporary barriers, and sandbagging incorporated as constraints within the system. This allowed comparison of how different mitigation strategies alter flood extent, depth, and infrastructure exposure.

Key components included:

• Constructing spatial models to simulate inundation under different environmental conditions
• Incorporating intervention scenarios by modifying terrain and flow constraints
• Analysing impact in terms of affected regions, infrastructure exposure, and changes in flood extent
• Conducting long-term cost analysis over a 50-year horizon, incorporating durability, labour, and damage estimates

The project integrates spatial modelling with economic reasoning to support decision-making under uncertainty. Results highlight that optimal mitigation strategies depend not only on initial effectiveness, but on long-term resilience under changing conditions.

Academic Misconduct Committee

University Governance | Evidence Evaluation | Institutional Decision-Making

Served as a student representative on the University of Melbourne Academic Misconduct Committee, participating in hearings relating to plagiarism, contract cheating, unauthorised collaboration, and misuse of AI tools.

The role involved reviewing sensitive and confidential case material, identifying inconsistencies in explanations, evaluating procedural fairness, and contributing to decisions with significant academic and disciplinary consequences. Cases often involved incomplete or conflicting information, requiring careful judgement under uncertainty rather than simple rule application.

Working within a confidential disciplinary process required maintaining discretion while handling sensitive personal information and evidence. The experience reinforced the importance of evidence-based reasoning, consistency of judgement, and understanding how individual behaviour interacts with institutional systems, incentives, and accountability structures.

Data Science Student Society & Responsible AI Development

Operational Coordination | AI Governance | Communication

Contributed to multiple student-led initiatives focused on industry engagement, AI systems, and responsible technology use through the Data Science Students Society (DSCubed) and Responsible AI Development.

Played a leading role in coordinating Industry Networking Night, managing communication between venue staff, student organisations, and external participants to organise logistics including catering, security, AV systems, seating, lighting, and event layout.

Also contributed to the DSCubed AI Projects Team, where work focused on improving AI-assisted onboarding and hiring processes used internally by the organisation. This included thinking about how AI systems could support operational workflows while remaining understandable and reliable for users.

Through Responsible AI Development and Green Impact initiatives, participated in discussions and evaluations surrounding responsible AI use, confidentiality risks, energy consumption, and broader social impacts of large-scale AI deployment.

Elevate Education

Adaptive Communication | Audience Engagement | Live Presentation Delivery

Delivered academic workshops and study-skills seminars to high school students across metropolitan and regional Victoria through Elevate Education. The role focused not only on presenting information, but on maintaining engagement and adapting delivery style dynamically across different classroom environments.

Presentations required continual adjustment based on audience behaviour, energy levels, and classroom dynamics. Sessions often involved balancing humour, authority, pacing, and audience participation in real time, particularly in environments where students were disengaged, fatigued, or difficult to involve. Delivery style, tone, and interaction patterns had to be adapted continuously depending on how different groups responded.

The role also required careful control of presentation mechanics such as voice projection, timing, room awareness, and conversational flow, particularly across classrooms with different acoustics, layouts, and group behaviour. Considerable preparation went into refining examples, delivery structure, and transitions to ensure material remained engaging, clear, and relatable to students from different backgrounds.

The experience reinforced the importance of behavioural awareness, adaptive communication, and maintaining clarity and engagement under unpredictable real-world conditions.